Subscribe, if this was forwarded to you

Hey there,

This week, we’re diving into some of the most exciting developments across the AI and tech landscape. Google’s new Gemini Omni model is redefining what’s possible with multimodal video generation, while Andrej Karpathy’s Neural Networks: Zero to Hero course is the perfect resource for anyone looking to master deep learning from scratch. We’re also spotlighting CodeGraph, the open-source tool that’s transforming AI-powered code exploration, and sharing a practical tip on using Hallmark to create truly unique AI-generated UIs. There’s plenty more in the news, from breakthroughs in language model memory to Tencent’s ultra-compact translation models.

Let’s dive in!

Cheers, Andreas & Sascha

In today's edition:

📚 Tutorial of the Week: Karpathy’s Neural Networks: Zero to Hero

🛠️ Tool Spotlight: CodeGraph: Fast AI code exploration tool

📰 Top News: Google launches Gemini Omni its superior multimodal video generation model

💡 Tip: Hallmark: Unique AI-generated UI designs

Let's get started!

Tutorial of the week

Neural Networks: Zero to Hero—Karpathy’s Free Course

If you’ve ever wanted to truly understand neural networks from the ground up, Andrej Karpathy’s “Neural Networks: Zero to Hero” is the must-see resource. This acclaimed, free video course walks you through building and training neural networks, starting from the basics and progressing to advanced architectures like GPT. With hands-on coding and practical exercises, it’s ideal for anyone looking to deepen their machine learning expertise.

• Covers everything from foundational neural network concepts to building modern language models and tokenizers, including GPT architectures.

• Each lecture features live coding, detailed explanations, and accompanying Jupyter notebooks for hands-on practice.

• Exercises and Colab notebooks are provided to reinforce learning and encourage experimentation.

• Accessible to learners with basic Python knowledge, but also valuable for experienced practitioners wanting to understand neural networks at a deeper level.

• The course is open-source, highly starred on GitHub, and continually updated with new lectures and resources.

Whether you’re a student, data scientist, or AI engineer, this course offers practical, in-depth learning that demystifies neural networks and empowers you to build your own models from scratch.

Tool of the week

CodeGraph — Semantic code graphs for blazing-fast AI code exploration

CodeGraph is an open-source tool that builds a semantic knowledge graph of your codebase, dramatically accelerating AI-powered code exploration and analysis. By pre-indexing symbol relationships, call graphs, and code structure, CodeGraph enables agents like Claude Code, Cursor, and Codex to answer architecture and trace questions with a fraction of the usual tool calls—cutting costs, latency, and token usage.

• Massive efficiency gains: Benchmarks across real-world codebases show CodeGraph delivers up to 92% fewer tool calls, 35% lower cost, and 46% faster responses for AI agents compared to traditional file scanning.

• Broad language and framework support: Out-of-the-box support for 20+ languages (TypeScript, Python, Go, Java, Swift, etc.) and deep framework-aware routing for popular web stacks (Django, Flask, Express, Rails, and more).

• Zero config, 100% local: No external services or API keys required—CodeGraph runs entirely on your machine, keeping your code private and secure.

• Seamless agent integration: One-command installer auto-configures leading AI coding agents (Claude Code, Cursor, Codex, Gemini, and others) to use the CodeGraph MCP server.

• Always up-to-date: Native file watchers (FSEvents/inotify/ReadDirectoryChangesW) keep the semantic graph in sync with your code as you work—no manual syncing needed.

CodeGraph is rapidly gaining traction among AI code tool users and has already been validated on large open-source projects. See the GitHub repo for star history and adoption trends.

Top News of the week

Google Unveils Gemini Omni: The “Anything-to-Anything” Model for Video and Beyond

Google has launched Gemini Omni, a groundbreaking multimodal AI model that can generate and edit videos from any combination of text, images, audio, and video inputs. Announced as the successor to Veo, Gemini Omni represents a major leap in AI’s creative and reasoning abilities, allowing users to create, transform, and fine-tune video content through natural, conversational prompts.

Gemini Omni stands out for its deep world understanding and native multimodality. Users can edit videos step-by-step—changing backgrounds, swapping objects, or adjusting styles—while maintaining scene consistency and real-world logic. The model leverages Gemini’s general knowledge, intuitive grasp of physics, and cultural context to produce photorealistic and meaningful outputs. Key features include multi-turn editing, reference-based style and motion transfer, native audio generation, and support for up to five photo references per video. All content is watermarked with SynthID for transparency and safety.

The rollout is available to Google AI Plus, Pro, and Ultra subscribers via the Gemini App and Google Flow, with free access for YouTube Shorts and YouTube Create users. Developers and enterprise customers will gain API access soon, signaling a new era for creative AI tools.

🔗 Read the official Gemini Omni announcement

New "Sleep" Method Boosts Language Model Memory and Reasoning

Researchers have introduced a "sleep" mechanism for large language models, inspired by biological memory consolidation. This approach allows models to periodically compress and organize long contexts into persistent memory without slowing down inference. Experiments show that increasing "sleep" duration significantly improves performance on tasks requiring deep reasoning over long or evicted contexts, such as multi-hop retrieval and math reasoning. The technique could pave the way for more scalable, enduring AI systems that handle complex, long-horizon tasks efficiently.

🔗 Read the research paper on arXiv

Figure Robots Expand to Major Retail Distribution Centers

Figure has signed a commercial agreement with Catalyst Brands to deploy its humanoid robots in the company’s Reno, Nevada distribution center. Catalyst operates well-known brands like JCPenney, Aéropostale, and Brooks Brothers. The rollout aims to automate physically demanding logistics tasks, enabling human workers to focus on higher-value activities. This partnership marks a significant milestone in scaling AI-powered robotics across multi-brand retail operations, reflecting growing momentum for automation in supply chain modernization.

🔗 Official announcement from Figure

New Study: Coding Agents Struggle with Real-World Backend Constraints

A comprehensive benchmark reveals that large language model (LLM) coding agents experience a sharp performance drop—averaging a 30-point decline in assertion pass rates—when tasked with backend code generation under strict architectural, database, and ORM constraints. While agents perform well with loose specifications, their reliability falls off in production-grade scenarios, especially with real databases and multi-file repositories. The main failure points are incorrect query logic and ORM runtime errors, highlighting the need for more robust, constraint-aware agent architectures.

🔗 Full paper: Constraint Decay—The Fragility of LLM Agents in Backend Code Generation

Tencent Releases Ultra-Compact, Open-Source Translation Models

Tencent has open-sourced the Hy-MT2 family of multilingual translation models, including a 1.8B-parameter version that can be compressed to just 440 MB using extreme quantization. The models support 33 languages and are designed for on-device deployment, offering high-quality translation without relying on cloud infrastructure. Benchmarks show that even the smallest model outperforms mainstream commercial APIs in many scenarios, making it a strong choice for privacy-sensitive or offline applications.

🔗 Explore Tencent Hy-MT2 on Hugging Face

Tip of the week

Supercharge AI-Generated UIs with Hallmark for Unique Designs

Tired of AI tools producing bland, cookie-cutter user interfaces? Hallmark is a free, open-source skill that helps AI agents like Claude Code, Cursor, and Codex generate visually distinctive, high-quality UIs.

• What it is: Hallmark acts as a rule set and design skill, guiding AI coding tools to avoid generic templates and produce more creative, differentiated interfaces. It runs outputs through 65 quality checks and offers 22 design themes.

• How to apply: Install Hallmark with npx skills add nutlope/hallmark, or copy the skill files into your tool’s directory. Use commands like hallmark audit <target> to score existing code, hallmark redesign <target> to rebuild with a new fingerprint, or hallmark study <screenshot | URL> to extract design DNA from admired sites.

• Why it's useful: Hallmark ensures every UI generated feels unique, not just a color-swapped template. It’s MIT licensed, easy to integrate, and works seamlessly with popular AI coding tools.

• Key benefit: Save time while elevating the quality and originality of your AI-generated interfaces—perfect for developers and designers seeking standout results.

Try Hallmark whenever you want your AI-generated UIs to stand out from the crowd.

Explore Hallmark and see live examples

We hope you liked our newsletter and you stay tuned for the next edition. If you need help with your AI tasks and implementations - let us know. We are happy to help